- Mathematics and Algorithm
- Deep Learning (DL): Proficient in Reinforcement Learning, Deeply understand DL optimization, experienced with PyTorch, well-informed with NLP (TransFormer, LSTM), Image Classification(ResNet, Inception), Recommendation System
- Machine Learning: Skilled in Supervised Learning (tree models such as GBDT, Logistic Regression), Unsupervised Learning (Clustering, Association Analysis), Semi-Supervised Learning, Contrastive Learning, probabilistic graphical model; expert on data analysis tools(Scikit-Learn, Pandas, Numpy, Sympy, Matplotlib, scipy)
- Data structures and algorithms, computational complexity
- Programming Language: Expert in Python, familiar with C/C++, bash, MATLAB, Makefile
- Linux: use ArchLinux/Ubuntu since 2016, managed 7 deep learning servers, own a VPS
- Software Engineering (git, CI, Dockerfile, pretty code style), Network (HTTP, OSI)
- Web Crawler: requests, bs4, selenium, developed Sukiya questionnaire auto filler
- Major and Profession
- Operations Research: MILP (ortools), Heuristic Optimization (GA/SA/CEM), Convex Optimization (cvxpy)
- Control Theory: Optimal Control, Linear Systems, System Identification
- Traffic Control and AutoPilot: Traffic Wave Theory, Webster Model, Green Wave Model, Path Planning
COMPANY POSITION TIME
Intelligent Transportation System (ITS): Research on Applying Reinforcement Learning (RL) to Real World Production
- S: Intelligent control of traffic light signal timing could enhance traffic efficiency greatly. Current state-of-the-art AI traffic control algorithms are mainly based on strong assumptions, which makes it hard to apply to real-world products.
- T: 1. Bring RL to real-world production, and make it outperform traditional algorithms; 2. Enhance RL training with expert experience; 3. Research on RL generalization among several traffic scenarios.
- A: 1. Design and develop a real-world compatible RL environment; 2. Implement traditional algorithms as baseline; 3. Fine-tune RL algorithm and environment; 4. Introduce information entropy to solve the problem of data imbalance, making RL agents learn expert experience and highly accelerating RL training. 5. Design an RL model which generalizes to any shape of traffic intersections.
- R: Published: PRIVATE; another paper about RL generalization is under review.
Vehicle to Everything (V2X): Cooperative Adaptive Cruise Control (CACC)
- S: CACC is a crucial application in autopilot from levels of L2 to L4. For example, there is a GLOSA defined in project C-ROADS (an EU project).
- T: Research and develop a CACC system to reduce energy consumption and improve driving comfort level.
- A: 1. Implemented energy consumption models (HBEFA, SUMO/Energy, VT-CPFM)；2. Recurrent GLOSA and Eco-CACC; 3. Designed and developed a V2X-CACC by using optimal control to perform vehicle longitudinal trajectory planning with consideration of vehicle to vehicle (V2V) and vehicle to infrastructure (V2I) interaction.
- R: For single CAV, energy consumption is reduced by 6% on the cost of a 1% travel time increment; while on 10% CAV penetration, the energy consumption is reduced by 9%.
ITS: Urban Traffic Signal Control System
- S: Urban traffic signal control system adjusts traffic light timing to ensure vehicles and pedestrians move safely and efficiently.
- T: Design and develop large-scale traffic signal control algorithm modules, and utilize machine learning algorithoms to improve efficiency and reduce congestion among several scenarios.
- A: 1. Scheme selection: Use clustering algorithm (DBSCAN & KMeans) to group traffic data with different characteristics, and use KNN to learn the mappings from traffic characteristics to signal plans; 2. Time-of-day division: divide the time of a day by sequence clustering algorithms (spectral clustering & Fisher); 3. Traffic Optimization: based on the traffic flow model, solve single intersection and arterial by Mixed-integer Linear Programming (MILP) modeling, solve congestion optimization by heuristic optimization modeling; 4. Reliability: embed traffic simulation tools (including SUMO) into the evaluation module with a fallback logic to shield poor signal plans.
- R: Published 3 SDKs including PRIVATE, and evaluated in PRIVATE that the traffic control system outperforms human experts.
- Obtained patents:
Reinforcement Learning-based Reactive Force field OptimizationTIME
- S: Reactive force field (ReaxFF) is a function with a large number of parameters, which usually takes several months to optimize when using traditional genetic algorithms.
- T: Explore a heuristic optimization method based on RL, to speed up the optimization process of the reaction field parameters of similar materials.
- A: Based on the assumption of similar materials having similar ReaxFF, design and implement the RL environment, perform feature engineering and implement different optimization methods (including RL and traditional genetic algorithms).
- R: 1. In terms of the optimization performance of similar materials, it has comparable results with the professional optimization software GARFField, greatly surpassing the conjugate gradient method and simulated annealing algorithm; 2. Obtain software copyright.
Data Mining-based Insurance Underwriting SystemCOMPANY TIME
- S: PRIVATE project with IBM, Baidu, etc to develop an automatic underwriting system to reduce human labor.
- T: Deeply understand customers' needs while communicating with underwriting experts, design and develop an automatic underwriting system, to win the bid.
- A: Perform feature engineering, design evaluation criterion, train and fine-tune Logistic Regression, Random Forest, LightGBM classification models; analyze bad cases and improve features; Encapsule the ML model into an SDK with C++.
- R: 1. Fulfilled PoC and won the bid; 2. Achieve 15% human labor reduction and 95% preception on the 530k training dataset.
Web Crawler and Website: ShadowSocksShareTIME
- S: Shadowsocks is a proxy software and was popular in China. Before the project was established, there were a lot of free accounts shared over the Internet but short of subscription services. It's not convenient for users without a subscription service.
- R: This project earns 350 sums up to ¥1871.91 donations and open source on GitHub with 3.1k Star and 1.1k Fork.
- Project link: ss.pythonic.life
- Strong: have exercise habits, periodically go hiking or riding
- Cooperative: 1. do research with interns on RL with expert experience and generalization; 2. communicate with product manager to decide product design, debug with the engineering team to make sure the urban traffic signal control system works well; 3. cooperate with MEC production line and RoboX to plan product layout together.
- Self-Learning and Fast Learner: gain deep mastery in website development, materials science, traffic engineering, and autopilot.
- Summarizing online notebook: notes.pythonic.life
- Geek Spirit: develop a bunch of scripts: CharlesScripts which earns 981 Star and 731 Fork; generated curriculum vitae by this tool; use Dvorak keyboard layout since 2018.
- Personal Blog: the0demiurge.blogspot.com
- Google Scholar: PRIVATE
- LinkedIn: linkedin.com/in/xuchengcheng